fix(cli): chunk oversized compactions

[catrielmuller]

Context

Fix CLI compaction overflow loops when long conversations or replayed turns still exceed the model context after the normal summary pass.

Implementation

• Add Kilo-specific chunked compaction that splits oversized histories, summarizes chunks with a smaller output budget, then synthesizes a final summary.
• Fall back to chunked compaction when the first compaction overflows, and compact oversized replay requests before auto-continuing.
• Cap LLM maxOutputTokens based on estimated input and tool schema tokens so small-context models keep room for prompts.

Screenshots

2026-05-14_11-09-17.mp4

How to Test

• From packages/opencode/, run bun test ./test/kilocode/session-compaction-chunks.test.ts.
• Exercise a long CLI session that triggers compaction; oversized histories should compact through chunk summaries instead of ending in a context overflow loop.

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Previously Flagged — All Resolved

File	Issue	Status
packages/opencode/src/kilocode/session/compaction-chunks.ts	WARNING: Concurrent mutation of input.agent.options — race condition	✅ Fixed: new agent const created, input.agent no longer mutated
packages/opencode/src/kilocode/session/compaction-chunks.ts	WARNING: removeMessage not in Effect.ensuring — orphaned messages possible	✅ Fixed: removeMessage now wrapped in Effect.ensuring, runs on both success and fiber interruption
packages/opencode/src/kilocode/session/compaction-chunks.ts	SUGGESTION: Depth counter increments twice on recovery path	✅ Addressed: synthesis call now uses depth + 2
packages/opencode/src/kilocode/session/llm.ts	SUGGESTION: !input.configured short-circuits on configured = 0	✅ Fixed: split into == null check and <= 0 check

Files Reviewed (2 files)

• packages/opencode/src/kilocode/session/compaction-chunks.ts — 0 issues
• packages/opencode/src/kilocode/session/llm.ts — 0 issues

Fix these issues in Kilo Cloud

---

_{Reviewed by claude-sonnet-4.6 · 217,950 tokens}